Health data helped identify kids in the ER who are likely to develop sepsis within 48 hours
Sepsis, or infection that causes life-threatening organ dysfunction, is a leading cause of death in children worldwide.
To prevent this rare but critical condition, scientists at Northwestern University and Ann & Robert H. Lurie Children’s Hospital of Chicago developed and validated AI models that accurately identify children at high risk for sepsis within 48 hours, so they can receive early preemptive care, as detailed in a study published in JAMA Pediatrics.
These predictive models used routine electronic health record (EHR) data from the first four hours the child spent in the emergency department before organ dysfunction was present.
The multi-center study is the first to use AI models to predict sepsis in children based on the new Phoenix Sepsis Criteria. Before this study, predictive models had not improved early diagnosis.
“The predictive models we developed are a huge step toward precision medicine for sepsis in children,” said corresponding author Elizabeth Alpern, MD, MSCE, vice chair and chief of Emergency Medicine in the Department of Pediatrics. “These models showed robust balance in identifying children in the emergency department who will later develop sepsis, without overidentifying those who are not at risk. This is very important because we want to avoid aggressive treatment for children who don’t need it.”
The study included five health systems contributing to the Pediatric Emergency Care Applied Research Network (PECARN), which provided Alpern and colleagues access to a large dataset and diverse population. The scientists discovered the machine-learning models by looking retrospectively at data from emergency department visits from January 2016 to February 2020. They then applied the models to data from 2021 to 2022 to validate how well the models performed. In both cases, they used data from the first four hours of care to try to predict what would happen in the next 48 hours of care.
Children with sepsis already at arrival or within the first hours of emergency-department care were excluded, focusing the goal of the study on predicting sepsis, to allow for early initiation of therapies that have been proven as lifesaving.
Predictive features included emergency department triage score, heart rate or respiratory rate, and already existing medical conditions like cancer.
“We evaluated our models to ensure that there were no biases,” said Alpern, who is also the George M. Eisenberg Professor of Pediatrics and division head of emergency medicine at Lurie Children’s. “Future research will need to combine EHR-based AI models with clinician judgment to make even better predictions.”
This project work was supported by the National Institute of Child Health and Human Development (NICHD) grant R01HD087363.
